Researchers Steal Reasoning Traces from Anthropic, OpenAI, and Google APIs via Replay AttackScaling Inherently Interpretable Language Models Challenges Capability-Interpretability Trade-offReasoning Models Fail to Allocate Test-Time Compute Efficiently Across QuestionsCMU-Drive and V2V-VLA: Cooperative Driving Benchmark and Vehicle-to-Vehicle VLA ModelsLorExperts and BTExperts: Shape-Mutating Compression for MoE ModelsDynamic Coalition Formation and Communication Pricing Framework for Agentic AI SystemsLUCID: Long-Horizon Humanoid Loco-Manipulation via Imagined DynamicsGoogle's AMIE Demonstrates Real-Time Clinical Video ConsultationsChai Discovery Closes 4 BioAI Deals This SummerDeepSeek Overtakes Google in Volume; Cost Per Token Drops 13.6%OpenAI Begins Testing Ads in ChatGPT to Support Free AccessOpenAI's Daybreak Cybersecurity Models Now Available on AWS BedrockMuse Glimmer and Spark: Open Weights Promise Personal Superintelligence on a Single RTX 3090Vercel Connect Adds Observability for Token Lifecycle TrackingVercel Enterprise Managed Users Now Generally AvailableVercel Connect Adds CLI Setup for 100+ IntegrationsGitHub Universe: AI Agents Shift Developers from Coders to OrchestratorsNo Lossless Text Transformations with LLMs: Simon Willison's AI Writing PolicyEverything Hackable Will Get Hacked: AI Reshapes Cyber Threats and DefenseClaude Code v2.1.228 Fixes Session Freeze and Git Bash DetectionResearchers Steal Reasoning Traces from Anthropic, OpenAI, and Google APIs via Replay Attack

A new arXiv paper ‘Scaling Inherently Interpretable Language Models’ challenges the premise that interpretability is a tax on capability. Instead of training opaque systems and explaining them post-hoc, the authors propose building inherently interpretable models. This research offers a new path for AI safety and transparency. Paper ID: arXiv:2608.07594v1.

New arXiv paper finds reasoning language models struggle to allocate test-time compute when multiple problems share cost or latency constraints. Models show significant degradation in cross-question scenarios, revealing a lack of global optimization in current reasoning strategies.

New paper presents CMU-Drive, a cooperative driving benchmark with reasoning annotations, and V2V-VLA, a vehicle-to-vehicle vision-language-action model enabling multi-vehicle perception fusion and coordinated action. Addresses the gap in cooperative driving for existing single-agent VLA models.

New paper introduces LorExperts and BTExperts, shape-mutating compression methods for MoE expert weight matrices. Unlike existing pruning/merging approaches that sacrifice accuracy, these methods maintain model quality while enabling cost-efficient deployment of large MoE models.

New paper proposes dynamic coalition formation and communication pricing for multi-agent AI systems. Fixed communication or full broadcast wastes tokens and latency. The framework lets agents form teams dynamically and price communication, cutting overhead while maintaining performance.

LUCID introduces a novel architecture for long-horizon humanoid loco-manipulation tasks. It uses a latent-skill unified control mechanism with imagined dynamics models, enabling robots to autonomously sequence whole-body skills (walking, grasping, carrying) and make high-level decisions without scripted planners or finite-state machines. Experiments show improved task success rates and smoother skill transitions over baselines. For robotics developers, this simplifies complex task decomposition and boosts deployment efficiency.

Google’s research medical AI system, AMIE, demonstrated real-time clinical video consultation capabilities in a first-of-its-kind study. In simulated settings, AMIE conducted live video consultations. This research showcases medical AI’s potential in telehealth and clinical assistance, though further validation is needed before real-world implementation.

The pharmaceutical industry is suddenly paying for Bio×AI tools, with Chai Discovery leading the pack after closing four deals this summer. Co-founder Matt McPartlon and Product leader Neil Patil explain why. This signals accelerating commercialization of AI in biopharma, with drug companies willing to pay for AI tools that deliver tangible value.

Vercel AI Gateway Production Index (Aug 2026): DeepSeek has overtaken Google in API call volume, while cost per token fell 13.6% month-over-month. Based on trillions of production tokens, the index reveals shifting enterprise AI usage patterns.

OpenAI announces it is beginning to test ads in ChatGPT to support free access. Ads will be clearly labeled, maintain answer independence, and include strong privacy protections and user controls. This marks a significant step in exploring sustainable free-tier models, though balancing user experience with commercialization remains key.

OpenAI and AWS bring Daybreak cybersecurity models to Amazon Bedrock, enabling enterprise security teams to leverage threat detection and vulnerability analysis capabilities directly within their AWS environment without data migration.

A win for open models: Muse releases Glimmer and Spark with open weights. Glimmer runs on a single RTX 3090, bringing near-superintelligent AI to individual developers. Spark targets higher performance tiers. Both models are available for download now.

Vercel Connect now provides line-level visibility into token lifecycles, showing who created tokens, which apps or projects used them, when, and their active status. Each connector’s detail page includes a new Observability tab with runtime events covering every token request, authorization, refresh, and revoke, addressing token transparency and auditability issues.

Vercel announces the general availability of Enterprise Managed Users (EMU). This feature gives organizations full control over Vercel accounts tied to verified domains, making the identity provider the single source of truth for authentication and account lifecycle. Accounts on company domains are centrally governed instead of managed separately, enhancing security and administrative efficiency.

Vercel Connect now supports CLI setup for all 100+ connectors. Previously only some services could be configured in the terminal; now passing the service name creates a connector and attaches it to your project, streamlining scripted workflows and CI/CD automation.

GitHub Blog argues that AI agents are shifting the developer’s role from pure coder to orchestrator of the delivery system around code. Developers now own more of this system, not just code itself. GitHub Universe will host discussions on this trend, allowing devs to connect, learn, and explore what’s next.

Developer Simon Willison shares his internal policy on acceptable AI writing use by engineers. The core argument: there are no lossless transformations of natural-language text. A short, high-quality read that supports its own recommendations. It reminds developers to be aware of potential information loss when using LLMs to polish writing and to maintain content quality control.

Vercel blog argues AI models’ cybersecurity capabilities have grown dramatically, reshaping both threats and defense. Defenders currently hold an advantage using stronger proprietary models, but attack tools are advancing rapidly, requiring continuous security strategy upgrades.

Claude Code releases version v2.1.228, fixing three issues: interactive sessions that could stop redrawing entirely while the process kept running after a rare internal layout error; git/Git Bash not being found on Windows when launched from a parent folder of the git installation; and /tui potentially reverting the session.